TSMC is nearly quadrupling CoWoS capacity to 130,000 wafers/month by late 2026 and, in a major strategic shift, outsourcing the complex front end CoW (Chip on Wafer) step to OSAT partners ASE, SPIL, and Amkor starting... Nvidia alone consumes about 60% of total CoWoS output, and CEO C.C.

Create a landscape editorial hero image for this Studio Global article: What are the key developments in TSMC's advanced packaging strategy, including its outsourcing of CoWoS AI chip packaging to partners like A. Article summary: TSMC is executing a multi-pronged advanced packaging strategy: aggressively scaling CoWoS capacity nearly 4x to 130,000 wafers/month by late 2026, outsourcing a significant share of back-end and even front-end CoW steps . Topic tags: general, general web, user generated, government. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, ch
TSMC is executing a multi-pronged advanced packaging strategy in 2026 that addresses the most pressing bottleneck in the AI chip supply chain. Faced with demand that has pushed CoWoS capacity to its limits, the company is scaling its own production nearly fourfold while outsourcing critical steps to OSAT partners and developing a new bridge-based packaging technology that directly competes with Intel's EMIB. The financial results are striking: TSMC's overall gross margin reached 67.7% in Q2 2026, and analysts believe it could approach 70% .
TSMC's CoWoS (Chip-on-Wafer-on-Substrate) packaging is the de facto standard for AI accelerators from Nvidia, AMD, and others, but capacity has been "extremely tight" and sold out through 2026, according to CEO C.C. Wei . Nvidia alone consumes about 60% of total CoWoS output, largely for its next-generation Rubin architecture
.
In response, TSMC is scaling its own CoWoS production from approximately 35,000 wafers per month in late 2024 to a target of 120,000–130,000 wafers per month by the end of 2026 — nearly a 4x increase . Bernstein expects total industry CoWoS capacity, including OSAT contributions, to reach about 1.25 million wafers per year
.
But the internal scaling alone is not enough. In a major strategic shift, TSMC is now outsourcing not just the simpler WoS (Wafer-on-Substrate) back-end step but also the more complex CoW (Chip-on-Wafer) front-end step to OSAT partners. This outsourcing to ASE, SPIL, and Amkor is expected to begin in the second half of 2026 . Volume estimates for 2026 suggest roughly 240,000–270,000 CoWoS wafers per year will move to OSAT partners — Amkor handling 180,000–190,000, SPIL 60,000–80,000, and ASE tripling its own capacity to 20,000–25,000 wafers per month by end-2026
. ASE has already raised advanced packaging prices 5–20% for 2026 in response to demand
.
JPMorgan estimates that from 2026 onward, roughly 70% of back-end steps for CoWoS-L (the highest-end variant used by Nvidia) will be outsourced to ASE and SPIL . The demand is being driven primarily by Nvidia's Rubin architecture and a surge in ASIC designs from hyperscalers
.
Intel's EMIB-T (Embedded Multi-die Interconnect Bridge) has gained significant traction. Google reportedly chose it for its 9th-gen TPU, and Intel's CFO claims the foundry is "close to closing some deals that are in the billions per year" on advanced packaging alone . TSMC's CoWoS constraints have pushed some customers to evaluate Intel's alternative
.
TSMC's response is an internally developed technology referred to as "EMIB-like" or "quasi-EMIB" . The company is partnering with Taiwan-based IC substrate maker Kinsus to create a bridge-based packaging solution that avoids the large silicon interposers used in CoWoS
. Intel's EMIB approach uses small embedded bridge chips only where chiplets need to connect, which lowers material cost and simplifies assembly
. TSMC's version is structurally similar
.
The project is still in development, but the move signals TSMC's recognition that Intel's EMIB is a "potent contender" and that a complementary or competitive bridge-based offering is needed to prevent customer loss . TrendForce notes that while TSMC retains a lead with its integrated turnkey solution and proven high yields, the growing AI market offers opportunities for both companies
. Intel's EMIB-T has reportedly hit 90% yield in technical verification — short of the 98% mass production standard, while TSMC already delivers above 98% on CoWoS packaging for AI accelerators
.
TSMC's aggressive packaging strategy is paying off financially. The company reported a gross margin of 66.2% in Q1 2026 and 67.7% in Q2 2026, well above the ~53–54% level a year earlier . Operating margin in Q1 was 58.1%
.
For Q2 2026, TSMC guided revenue of $39.0–$40.2 billion and gross margin of 65.5%–67.5% . Actual Q2 revenue came in at $40.2 billion, above the high end
. For Q3 2026, TSMC projects revenue of $44.6–$45.8 billion, representing about 12% sequential growth and 37% year-over-year growth at the midpoint
.
Analysts expected approximately $40 billion in Q2 2026 revenue (roughly 32% YoY growth) with EPS up over 50% YoY . TrendForce reported that some analysts believed gross margin could approach 70% in Q2 2026 given strong 3nm/5nm AI demand
. The Q2 results confirmed that trajectory.
Bernstein raised its TSMC target price citing "more CoWoS, more XPU & 23% revenue growth in 2026," with CoWoS becoming an increasingly meaningful revenue contributor . Advanced packaging contributed approximately 8% of TSMC's total revenue in 2025, with expectations to surpass 10% in 2026
.
The key tension in the strategy is that TSMC's aggressive outsourcing relieves a critical capacity bottleneck but shifts some high-value packaging revenue to OSAT partners. This protects customer relationships and maintains AI chip supply momentum, while the company's core foundry pricing power — especially on 3nm and 5nm nodes — continues to drive overall margins toward 70%.
Studio Global AI
Use this topic as a starting point for a fresh source-backed answer, then compare citations before you share it.
TSMC is nearly quadrupling CoWoS capacity to 130,000 wafers/month by late 2026 and, in a major strategic shift, outsourcing the complex front end CoW (Chip on Wafer) step to OSAT partners ASE, SPIL, and Amkor starting...
TSMC is nearly quadrupling CoWoS capacity to 130,000 wafers/month by late 2026 and, in a major strategic shift, outsourcing the complex front end CoW (Chip on Wafer) step to OSAT partners ASE, SPIL, and Amkor starting... Nvidia alone consumes about 60% of total CoWoS output, and CEO C.C. Wei has acknowledged the capacity is sold out through 2026.
JPMorgan estimates that from 2026 onward, roughly 70% of back end steps for CoWoS L will be outsourced to ASE and SPIL.